v0.1.7

vectorize-io/hindsightv0.1.7Jan 16, 2026by chonknick

Upgrading?

Release notes summary

Version 0.1.7 introduces support for the Voyage 4 family of embedding models, including flagship, mid-sized, lightweight, and open-weight variants. All models support 32,000 token context, flexible Matryoshka dimensions, and multiple quantization formats.

Key Highlights

  • Added voyage-4-large, the flagship MoE model ranked #1 on MTEB (78.26)
  • Added voyage-4, voyage-4-lite, and open-weight voyage-4-nano (Apache 2.0)
  • All Voyage 4 models support 32,000 token context
  • Flexible dimensions (2048/1024/512/256) via Matryoshka learning
  • Multiple quantization formats supported (float, int8, binary)

New Features

  • Voyage 4 embedding model support (voyage-4-large, voyage-4, voyage-4-lite, voyage-4-nano)
  • 32,000 token context window for Voyage 4 models
  • Matryoshka flexible dimension support
  • Float, int8, and binary quantization formats

Full Release Notes

## What's New

- Added support for Voyage 4 embedding models:
  - `voyage-4-large` - Flagship MoE model, #1 on MTEB (78.26)
  - `voyage-4` - Mid-sized balanced model
  - `voyage-4-lite` - Lightweight model, #9 on MTEB (73.29)
  - `voyage-4-nano` - Open-weight model (Apache 2.0)

All Voyage 4 models support:
- 32,000 token context
- Flexible dimensions (2048/1024/512/256) via Matryoshka learning
- Multiple quantization formats (float, int8, binary)

Related Repositories